Senate mulls offensive AI, new training tools and now Chinese faceswaps Trump
It's the wacky week in AI
Roundup Your weekly dose of tidbits from the AI world, beyond everything we've already covered, begins with a senate committee hearing in which a US lieutenant general, currently a nominee for the role of the director of the NSA, spoke about his concerns around the technology.
And it ends with a CEO of a Chinese AI startup demonstrating how AI can be used to perform a faceswap on Trump and Obama.
Lieutenant General Paul Nakasone, currently the commander of the United States Army Cyber Command, was quizzed by Senator Ted Cruz (R-TX) about his thoughts on AI.
The US Senate Committee on Armed Services was considering the nomination of Nakasone for the role of director of the NSA, as well as Dr Brent Park to be deputy administrator for Defense Nuclear Nonproliferation for the National Nuclear Security Administration, and Anne White to be assistant secretary of energy for environmental management for the department of energy.
Senator Cruz brought up the idea of poisoning systems with adversarial examples, something that was discussed during the first congressional hearing on AI he chaired last year.
Cyberterrorism in the future will not occur by DDoSing or “bringing the system down”, Cruz said. “But far more subtly, simply changing the data in the big data datasets so that the AI algorithms reach the wrong results”.
Adversarial machine learning is a big area of study in research, which has shown that these systems are shockingly brittle. Fuzzing a few pixels here and there can fool convolutional neural nets.
Nakasone agreed and called data “the coin of the realm”. He didn’t really address the problem of adversarial examples, but did talk about the need to verify any changes to the organization’s code.
"So Senator, previously we thought of only securing out networks. And what we’ve certainly learned is the fact that securing our data, which I’d say is ‘the coin of the realm,’" he said.
"Our data is critical. Think of the dangers that are posed of our data is manipulated, whether or not it’s in our financial, our health, our national defence records - it’s very, very critical for what we are doing. But also think of the security for our weapon systems that go with it. The code that underlines our platforms, the code that underlines the critical capabilities that our army, navy, airforce, and marines rely on."
"In terms of what must be done, I would offer that we have to think more broadly on term of defense and depth strategies as we look to the future. You highlighted the challenges of AI. Just as critical as AI might be for a terrorist, it’s critical for us to verify code."
"To be able to have the capability to verify the integrity of our data. And so I do see this as one of the areas that both has tremendous positive impacts for our nation and one that we must be able to understand the limitations and the consequences as well."
Another interesting question Cruz asked the lieutenant general was about how the NSA can compete with Silicon Valley for AI talent. There aren’t that many devs around with a specialized background in machine learning to fill the many roles created by the current hype.
Major tech giants boast six figure salaries, a share in its stocks, and rich resources, so how can the NSA snare people away?
Nakasone said that “no one in the private sector has the responsibility for defending the nation. We have a very, very unique mission and I think that resonates with young people today.”
“But I also think we have to have an approach to recruiting people that is dynamic, that tries different ideas and has unique partnerships - that is able to leverage ideas that may not be traditional within out military sphere. And I think that’s important because the space that we operate in is changing every single day and so why should our ideas change just as rapidly?"
Tension in Tensorflow
TensorFlow, the most commonly used open source machine learning framework, has been updated.
Its GitHub page lists the bugs that have been fixed, new features, and some major changes. It includes “prebuilt binaries are now built against CUDA 9.0 and cuDNN 7 and prebuilt binaries will use AVX instructions” - this may break TensorFlow on older CPUs, developers warned.
You can play around with it here.
Allen and AI
Paul Allen, Microsoft’s co-founder, announced a $125m fund for the Allen Institute for Artificial Intelligence (AI2) to kickstart a project to teach machines “common sense”.
It’s tricky to define what that exactly means. AI2 have defined it as “the everyday knowledge that virtually every person has but no machine does”.
That’s a pretty broad statement, so we asked AI2 for more information. A spokesperson explained to us: “Common sense for machines is the background knowledge that would allow an AI system to successfully accomplish general, non-specific tasks, something no AI today is truly capable of.
“We want to provide systems with the framework to understand nuances and likely outcomes of real-world scenarios, making AI much more robust and flexible to tasks that involve unstructured problem-solving or managing unanticipated situations. Current systems are very brittle; they can fail with the slightest change to the expected parameters of the task they are performing.”
Project Alexandria will be led by CEO Oren Etzioni. It will also build upon previous work from Project Aristo, concerned with machine reading and reasoning; Project Euclid that looks at question answering in mathematics for natural language understanding; and Project Plato, a challenge to understand images and visual scenarios.
“In AI literature, broarder common-sense type challenges are sometimes referred to as "multi-task learning" or "transfer learning" and "zero-shot" or "one shot" learning; but with the exception of word embeddings, modern AI has struggled to create background knowledge that yields a substantial boost across a wide range of tasks.
“One of our goals for Project Alexandria is to drive the development of common sense in the field of AI by introducing standard measurements for the common sense abilities of an AI system, which will give researchers a way to test current systems and inspiration for designing new ones,” the spokesperson added.
Train your own classifier
This week Microsoft announced new computer vision APIs from Cognitive Services on its cloud platform, Azure.
The Custom Vision service on its Azure Portal will make it easier for developers to train image classifiers to recognise objects they’re interested in. Retailers might want to train models to look out for certain clothing items or astronomers may want to find pictures of certain planets, etc.
The service can also be used on mobile phones. Models can be exported to CoreML for iOS11 and TensorFlow for Android devices. Its Face API also from its Cognitive Services and available on Azure has been expanded to recognize up to a million faces.
Microsoft called it “million-scale recognition capabilities”. It represents a new type of “person group” containing up to a million people, and a new type of face list with up to a million faces.
All of these images can now be sorted using Bing Entity search, a service that has been integrated with Azure Portal.
Free ML crash course
Are you interested in learning the basics of machine learning? Well, here’s a free short course from Google.
It’s aimed at people with no background of machine learning. But to make some sense of it it’s recommended that you scrub up on some algebra and have some programming proficiency, and in particular Python coding skills.
It starts off introducing some basic concepts such as logistic regression, classification, neural networks and moves onto ML engineering, and some real world applications. Each lesson is based on a video tutorial, notes and a quiz at the end.
You can start here.
MIT Intelligence Quest
SenseTime, a Chinese computer vision startup is the first company to join the MIT Intelligence Quest, an institute interested in applying machine intelligence to a range of interdisciplinary areas.
Xiao’ou Tang, who founded SenseTime is also an alumnus of MIT. The alliance between both groups “aims to open up new avenues of discovery across MIT in areas such as computer vision, human-intelligence-inspired algorithms, medical imaging, and robotics; drive technological breakthroughs in AI that have the potential to confront some of the world’s greatest challenges,” according to a blog post.
The startup is valued at more than $3bn. During the launch for the initiative, Tang took to the stage to demonstrate some of the current capabilities in computer vision. Jack Clark, strategy and communications director at OpenAI, filmed a short video where a clip of Obama could be superimposed to appear as Trump.
Here is the founder of SenseTime, Xiao'ou Tang, showing off a Trump<>Obama generative AI faceswap at launch for an MIT AI initiative that SenseTime is helping to fund. 2018 is delivering in terms of AI-geopolitics weirdness! https://t.co/emK9laObfu pic.twitter.com/iTHy5lVmwJ— Jack Clark (@jackclarkSF) March 1, 2018
It’s not too dissimilar from the whole Deepfakes pornography kerfuffle. Eep. ®
Sponsored: What next after Netezza?